dim_action = 40
dim_state = 3
am = ActionMap(dim_action)
agent = DQNAgent(dim_state, dim_action, am)
high = np.array([1., 1., 8.])
low = -high
#print("mean 100 episode reward before learning: {}".format(calculate_mean_reward(agent, env)))
episodes = 1000
for i in range(episodes):
print(i)
observation = env.reset()
while True:
env.render(mode="human")
action = agent.act(observation)
new_observation, reward, done, info = env.step(
env.action_space.sample())
agent.remember(observation, action, reward, new_observation, done)
if done:
agent.replay(100)
break
print("mean 100 episode reward after learning: {}".format(
calculate_mean_reward(agent, env)))
def run(env: RobotArmEnvironment,
agent: DQNAgent,
num_episodes: int,
max_num_steps: int,
batch_size: int,
directory_path: str,
random_run: bool = False):
reward_history_file_name = directory_path + "reward.csv"
action_history_file_name = directory_path + "action.csv"
max_q_history_file_name = directory_path + "max-q.csv"
state_history_file_name = directory_path + "state.csv"
# Parse these files with:
# with open(file_name, "r") as f:
# reader = csv.reader(f, delimiter=" ")
# for row in reader:
# for col in row:
# col = ast.literal_eval(col) # (nan values have to be checked for)
previous = time.time()
for episode_idx in range(num_episodes):
state = env.reset()
for step_idx in range(max_num_steps):
with open(state_history_file_name, "a") as f:
f.write(
("(" + ("{}," * 6) + ") ").format(*env.simulation.state))
if episode_idx % 100 == 0:
env.render()
time.sleep(1 / 10)
# take an action
if random_run:
action = env.action_space.sample()
else:
max_q, action, prediction = agent.act(state)
if not random_run:
with open(max_q_history_file_name, "a") as f:
f.write("{} ".format(max_q))
with open(action_history_file_name, "a") as f:
f.write("({},{}) ".format(
env.action_map.get(int(action))[0],
env.action_map.get(int(action))[1]))
# observe effect of action and remember
new_state, reward, done, info = env.step(action)
if not random_run:
agent.remember(state, action, reward, new_state, done)
with open(reward_history_file_name, "a") as f:
f.write("{} ".format(float(reward)))
# store new state
state = new_state
if done:
break
if not random_run:
agent.replay(batch_size)
# new line in all data files
with open(action_history_file_name, "a") as f:
f.write("\n")
with open(reward_history_file_name, "a") as f:
f.write("\n")
if not random_run:
with open(max_q_history_file_name, "a") as f:
f.write("\n")
with open(state_history_file_name, "a") as f:
f.write(("(" + ("{}," * 6) + ") \n").format(*env.simulation.state))
if not random_run and episode_idx % 50 == 0:
agent.save(directory_path + "weights-ep-{}".format(episode_idx))
current = time.time()
print("{}: episode {:3}/{:3} completed in {:4}s".format(
os.getpid(), episode_idx, num_episodes, current - previous))
previous = current
ct = time()
total_time_acting = 0
total_time_stepping = 0
total_time_remembering = 0
total_overhead = time()
ct_act, ct_step, ct_rem = 0, 0, 0
for i in range(max_iterations_per_episode):
if (episode_idx + 1) % 50 == 0 or episode_idx == 0:
env.render()
# sleep(1 / 2)
ct_act = time()
action = agent.act(state)
total_time_acting += time() - ct_act
ct_step = time()
next_state, reward, done, _ = env.step(action)
# print("action {}, state {}".format(env.action_map.get(action), next_state))
total_time_stepping += time() - ct_step
ct_rem = time()
agent.remember(state, action, reward, next_state, done)
total_time_remembering += time() - ct_rem
state = next_state
tr += reward
agent.replay(32)
def run_experiments(reward_index):
if reward_index < 0 or reward_index > 1:
raise ValueError()
num_episodes = number_of_episodes
num_steps = max_iterations_per_episode
batchsize = 32
state_size = 6
action_size = 81
memory_size = 100000
epsilon_start = 1
epsilon_min = 0.1
epsilon_decay_per_step = 10000
lr = 0.00001
dr = 0.99
amount_layers = 2
amount_nodes_layer = 40
frequency_updates = 1000
parameters = {}
parameters['num_episodes'] = num_episodes
parameters['num_steps'] = num_steps
parameters['batchsize'] = batchsize
parameters['state_size'] = state_size
parameters['action_size'] = action_size
parameters['memory_size'] = memory_size
parameters['epsilon_start'] = epsilon_start
parameters['epsilon_min'] = epsilon_min
parameters['epsilon_decay_episodes_required'] = epsilon_decay_per_step
parameters['learning_rate'] = lr
parameters['discount_rate'] = dr
parameters['amount_layers'] = amount_layers
parameters['amount_nodes_layer'] = amount_nodes_layer
parameters['frequency_update_target_model'] = frequency_updates
agent = DQNAgent(6, 81, num_steps, epsilon_start, epsilon_min,
epsilon_decay_per_step, dr, lr, amount_layers,
(amount_nodes_layer, amount_nodes_layer),
frequency_updates)
with RobotArmEnvironment(reward_function_index=reward_index,
reward_function_params=(1 / 6 * np.pi, 2 * np.pi,
1, 10, 0.05, 0.1, 2,
0.001, 1)) as env:
ah = list()
rh = list()
for episode_idx in range(number_of_episodes):
state = env.reset()
tr = 0
ct = time.time()
ah.append(list())
rh.append(list())
for i in range(max_iterations_per_episode):
action = agent.act(state)
ah[episode_idx].append(env.action_map.get(int(action)))
next_state, reward, done, _ = env.step(action)
rh[episode_idx].append(float(reward))
agent.remember(state, action, reward, next_state, done)
state = next_state
tr += reward
if done:
break
agent.replay(32)
print(
"episode {}/{}, average reward {}, epsilon {}, time taken {}s".
format(episode_idx + 1, number_of_episodes, tr,
agent.get_epsilon(),
time.time() - ct))
agent._update_epsilon()
if episode_idx % 100 == 0 and episode_idx != 0:
agent.safe()
save_info(episode_idx, reward_index, parameters,
env.action_map.to_json_object(), rh, ah,
env.to_json_object())